Flow Matching — The Simulation-Free Recipe Under Modern Diffusion

Paper. Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, Matt Le. Flow Matching for Generative Modeling. ICLR 2023. Meta AI · FAIR. [arXiv] · [OpenReview]

📑 Companion slide deck. A self-contained HTML walk-through of this review (in Korean) is available at Flow Matching 논문 해설 (standalone).html. The deck and this post share the same structure; the post is the long-form English version.

0. The Picture in One Paragraph

Continuous Normalizing Flows (CNFs) are an elegant generative model — sample $x_0$ from a simple prior, integrate an ODE $\dot{x} = v_\theta(t, x)$, get a sample from a complicated target. They’ve been theoretically attractive for years and practically unusable, because the standard training objective (maximum likelihood) requires solving the ODE forward and backward on every training step. That’s the “simulation cost.” Flow Matching proposes a different objective: instead of training to maximize likelihood, directly regress the vector field $v_\theta$ against a target vector field that produces a chosen probability path. The catch is that the marginal target vector field is intractable. The contribution is showing that you can construct it as a mixture of per-sample conditional vector fields, and the conditional flow matching (CFM) loss — which only ever sees the conditional field — has the same gradient as the intractable marginal loss. This eliminates simulation entirely. With optimal-transport conditional paths (straight-line interpolation between noise and data), CNFs trained this way outperform DDPM on CIFAR-10 and ImageNet on FID, NLL, and NFE simultaneously — the first time those three knobs moved together. The recipe is now the training objective under Stable Diffusion 3, Wan, and DreamZero.

1. The Background — Why CNFs Were Stuck

A CNF defines a generative model via a learned vector field $v_\theta: [0,1] \times \mathbb{R}^d \to \mathbb{R}^d$ acting through the ODE

\[\frac{dx(t)}{dt} = v_\theta(t, x(t)), \qquad x(0) \sim p_0.\]

The flow $\phi_t$ pushes the prior $p_0$ forward to a model distribution $p_t$. Training maximizes the log-likelihood under the change-of-variables formula:

\[\log p_1(x) = \log p_0(\phi_1^{-1}(x)) - \int_0^1 \nabla \cdot v_\theta(t, \phi_t^{-1}(x)) \, dt.\]

This is gorgeous — exact likelihood, single network — and expensive: each gradient step needs an ODE solve for every sample, and gradients flow through the solver. Scaling CNFs to ImageNet was a known open problem precisely because of this simulation tax.

2. The Move — Regress the Vector Field Directly

Flow Matching’s reframing is to give up on maximum likelihood as the training signal and instead specify a target by saying what probability path $p_t$ you want the model to follow, then regress the model’s vector field against the field that produces that path.

2.1 The (intractable) Flow Matching loss

Fix a probability path $p_t$ that interpolates $p_0$ (noise) to $p_1$ (data). Suppose we know the vector field $u_t$ that generates $p_t$ (in the sense that pushing $p_0$ along $u_t$ yields $p_t$). Then the natural loss is

\[\mathcal{L}_{\text{FM}}(\theta) = \mathbb{E}_{t, x \sim p_t} \big\| v_\theta(t, x) - u_t(x) \big\|^2.\]

A clean MSE between two vector fields. Simulation-free: no ODE solve, no divergence integral. If we had $u_t$, we’d be done.

The problem: $u_t$ is intractable for non-trivial paths. We can’t write it down in closed form for the marginal path; that’s exactly the thing the model is supposed to learn.

2.2 The CFM trick — replace the marginal with the conditional

Here’s the move that makes everything work. Construct $p_t$ as a mixture of conditional paths:

\[p_t(x) = \int p_t(x \mid x_1) \, q(x_1) \, dx_1,\]

where $x_1 \sim q$ is a data sample and $p_t(\cdot \mid x_1)$ is a simple conditional path from the prior to a point mass (or near-point-mass) at $x_1$. For the conditional path, the generating vector field $u_t(x \mid x_1)$ is tractable in closed form — it’s just the field that smoothly transports prior samples to $x_1$.

Now define the Conditional Flow Matching loss:

\[\mathcal{L}_{\text{CFM}}(\theta) = \mathbb{E}_{t, x_1 \sim q, x \sim p_t(\cdot \mid x_1)} \big\| v_\theta(t, x) - u_t(x \mid x_1) \big\|^2.\]

The model now regresses against a conditional field — which we have a closed form for — instead of the marginal field — which we don’t.

2.3 The theorem that makes this work

The paper’s central technical result:

$\nabla_\theta \mathcal{L}{\text{FM}}(\theta) = \nabla\theta \mathcal{L}_{\text{CFM}}(\theta).$

The two losses are not equal (they differ by a constant that depends only on the data and the path), but their gradients are identical. Optimizing the tractable CFM loss is equivalent to optimizing the intractable FM loss.

The proof uses the fact that the marginal vector field $u_t(x)$ is the posterior expectation of the conditional fields:

\[u_t(x) = \mathbb{E}_{q(x_1 \mid x_t = x)} \big[ u_t(x \mid x_1) \big],\]

i.e. the marginal field at $x$ is a weighted average of the conditional fields toward each possible data point, weighted by how likely each data point is to have produced $x$ at time $t$. The MSE regression to a marginal mean is equivalent (in gradient) to regression to the per-sample target.

This is the same structure that underlies denoising score matching — and the connection is not accidental. Flow Matching generalizes score matching: with a specific choice of conditional path (the variance-exploding or variance-preserving diffusion path) and a specific reparameterization, CFM reduces to the standard score-matching loss. Diffusion is one instance of Flow Matching, not the other way around.

3. Optimal-Transport Conditional Paths

Now the design freedom: which conditional path do we use? FM works for any path, but the paper makes a specific recommendation that turns out to dominate.

3.1 The OT conditional path

Pick the conditional path that linearly interpolates between a prior sample and a data sample:

\[\mu_t(x_1) = t \, x_1, \qquad \sigma_t(x_1) = 1 - (1 - \sigma_{\min}) t,\]

i.e. $x \sim \mathcal{N}(t x_1, (1 - (1-\sigma_{\min}) t)^2 I)$. At $t=0$, you’re at the standard Gaussian prior; at $t=1$, you’re at $x_1$ with a tiny residual variance $\sigma_{\min}$.

The associated conditional vector field is

\[u_t(x \mid x_1) = \frac{x_1 - (1 - \sigma_{\min}) x}{1 - (1 - \sigma_{\min}) t},\]

which — at $\sigma_{\min} \to 0$ — is the constant velocity $x_1 - x_0$ along a straight line from prior sample $x_0$ to data sample $x_1$.

This is the displacement interpolant from optimal transport, restricted to per-sample conditionals.

3.2 Why straight lines win

Diffusion conditional paths are curved — they push the sample toward $x_1$ along a noise-scaled trajectory that bends through high-noise regions before settling. Curved trajectories require more ODE-solver steps to integrate accurately at inference time.

OT paths are straight. A straight integral curve can be integrated with very few steps — in the limit of an ideal model, one step. Empirically the gain is dramatic: same FID, fraction of the function evaluations.

3.3 What FM-OT inherits

The OT path keeps every good property of the diffusion path while shortening the integration:

Closed-form conditional field (no learning $u_t$).
Closed-form sampling from $p_t(x \mid x_1)$ (just a Gaussian).
Straight integration → low NFE.

Nothing about the framework changes when you swap the conditional path; only the per-sample $u_t$ formula changes. The CFM gradient- equivalence theorem still holds.

4. Headline Results

All numbers are training the same U-Net architecture with the same budget on the same hyperparameters; the only variable is the training objective. Lower is better on all three metrics.

4.1 CIFAR-10

Method	NLL ↓	FID ↓	NFE ↓
DDPM	3.12	7.48	274
FM (diffusion path)	3.10	8.06	183
FM-OT	2.99	6.35	142

4.2 ImageNet 32×32

Method	NLL ↓	FID ↓	NFE ↓
DDPM	3.54	6.99	262
FM-OT	3.53	5.02	122

4.3 ImageNet 64×64

Method	NLL ↓	FID ↓	NFE ↓
DDPM	3.32	17.36	264
FM-OT	3.31	14.45	138

4.4 ImageNet 128×128

FM-OT: NLL 2.90, FID 20.9 — first time a CNF result is on this scale.

4.5 What’s striking about the table

Generative-model literature typically improves one of (FID, NLL, NFE) by hurting the others — better quality usually costs more function evaluations; better likelihood usually costs visual quality. FM-OT improves all three at once on every dataset. That kind of clean dominance is rare and is the reason this paper landed as hard as it did.

5. What the Method Actually Says

Stripped of acronyms, the structural claim:

Continuous Normalizing Flows are a useful generative model class, and the obstacle to using them was the simulation cost of maximum-likelihood training.

You can replace likelihood with vector-field regression. The marginal target is intractable, but if you express the target path as a mixture of per-sample conditional paths, the conditional regression has the same gradient as the marginal regression.

The choice of conditional path is a design knob. Straight-line (optimal-transport) paths are dramatically more efficient at inference time than the curved paths implicit in diffusion training.

Diffusion training is one instance of this framework, not the other way around. The framework gives you a strictly larger design space, and within it FM-OT is a sweet spot.

This is the kind of result that doesn’t just improve a benchmark — it reorganizes the field around a different abstraction.

6. Why It Matters — Three Years Later

Reading this paper in 2026, with the benefit of hindsight, three things stand out:

It became the training objective under most modern generative models. Stable Diffusion 3, Wan, Cosmos, and the DreamZero backbone all train with flow matching (often with OT paths) rather than the original score-matching diffusion objective. What looked like a clever theoretical paper in 2022 became the default by 2024.
The CFM trick is reusable. Any time you want to regress a network against an intractable marginal target that’s a mixture of tractable conditionals — molecule generation, video synthesis, robot action chunks — the same gradient-equivalence argument applies. The paper effectively opened a generic recipe.
It exposed a knob the field didn’t know it had. Before this paper, “what probability path do we use?” wasn’t a question you saw in diffusion papers — the variance-preserving and variance-exploding paths were treated as given. Flow Matching made the path itself a design choice, which is what unlocked straight-line OT paths and, later, more exotic choices (rectified flows, consistency flow matching, etc.).

7. Limitations Worth Knowing

Theory is asymptotic. The gradient-equivalence theorem holds in expectation over infinitely many samples. At finite batch size, CFM has higher variance than directly fitting the marginal field would (if we could). In practice this is fine, but it’s worth knowing the theorem is not a free lunch.
OT path is displacement OT, not a true Monge optimal map. The “optimal transport” path is per-sample straight-line interpolation between a noise sample and a data sample — which noise sample pairs with which data sample is random. True OT would pair them to minimize transport cost, which the original paper does not do. The follow-up minibatch OT work (Tong et al., 2023) closes that gap.
No conditional generation in the original. The paper covers unconditional generative modeling. Class- or text-conditioned FM required follow-up engineering; Stable Diffusion 3 and Wan are the modern conditional implementations.
Architecture not the contribution. The U-Net used is the standard one from the diffusion literature. The contribution is the training objective; the architecture is a control variable. This is also why the recipe transferred so cleanly to other backbones (DiT, transformer-based video models) later.

8. The Takeaway for a First Reader

If you remember three things:

CNFs are simulation-free trainable if you regress the model’s vector field against a target field instead of maximizing likelihood. The target field is intractable in the marginal but tractable in the conditional, and the conditional MSE has the same gradient as the marginal MSE.
The conditional path is a design choice. OT (straight-line) paths dominate diffusion paths because straight curves integrate cheaply at inference time.
The result is the first generative training recipe that wins on FID, NLL, and NFE simultaneously on CIFAR-10 and ImageNet, and the recipe became the default training objective under Stable Diffusion 3, Wan, and modern video / world models.

That’s the arc: maximum-likelihood CNFs were stuck → reframe training as vector-field regression → use the conditional trick to make the regression tractable → use OT paths to make inference cheap → end up with the substrate under most modern generative modeling.

References

Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., & Le, M. (2023). Flow Matching for Generative Modeling. ICLR 2023. arXiv:2210.02747.
Tong, A. et al. (2023). Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv:2302.00482.
Background: DDPM (Ho et al., 2020); Score-based generative modeling (Song et al., 2021).
Related on this site: DreamZero paper review — the world-action-model backbone that depends on this objective; Parcae paper review — for the discretization tricks Parcae borrows from SSMs and FM shares structural family with.